Professional Documents
Culture Documents
Hopgood
(Guest Editors)
Abstract
In this paper, we present a representation for three-dimensional geometric animation sequences. Different from
standard key-frame techniques, this approach is based on the determination of principal animation components
and decouples the animation from the underlying geometry. The new representation supports progressive animation compression with spatial, as well as temporal, level-of-detail and high compression ratios. The distinction of animation and geometry allows for mapping animations onto other objects.
Keywords: Geometric animations, polyhedral morphing, multiresolution/progressive representation
1. Introduction
Computer animation has evolved into a standard technique
in Computer Graphics. In the last few decades, a number of
different animation techniques have been developed, but
key-frame animation has established itself as the standard
technique for describing time-dependent 3D Computer Animation. Instead of describing every single frame, only a
sequence of principal frames - so called key-frames - is
defined and additional frames are generated by interpolating between two consecutive key-frames using in-betweening.
Elaborated techniques have been developed to allow for
the automated generation of physically-based behavior,
namely kinematics and inverse kinematics, and today animation systems are increasingly coupled with simulation
engines to facilitate the production of complex and realistic
animations. However, key-frames remain the standard for
representing animations.
Despite its widespread use, this concept of merging
object and geometry descriptions has a fundamental problem. While it is easy to specify, design, or output an animation in terms of key-frames, it is difficult to manage due to
the large amount of data or to change the time-behavior
since all meta-information is lost. In particular, the following problems connected to geometric key-frame animations
can be stated:
Redundancy: A complete object description has to be
recorded for each key-frame, even if parts of the object
do not change at all. Additionally, repetitive patterns
The Eurographics Association and Blackwell Publishers 2000. Published by Blackwell
Publishers, 108 Cowley Road, Oxford OX4 1JF, UK and 350 Main Street, Malden, MA
02148, USA.
result in repeating the geometry description. Consequently, animation sequences are usually very large
and hard to apply in streaming applications. The compression of such sequences is a problem which is
under intense investigation in the Computer Graphics
research community today.
Modification: Exchanging an animated object in a
scene while reusing the once specified animation at the
same time is an involved task, even though most of the
necessary information should be available. In general,
after introducing the new model, it has to be animated
again by hand. Similarly, it is almost impossible to
make use of a once defined animation and to extend or
exchange the objects behavior.
The aspect of reuse has been addressed for specific
object domains. One approach is standardized parameterization of an object and its possible behaviors, thus
allowing for the exchange of both geometry and animation (e.g. Humanoid Animation26 or Facial
Animation20 in MPEG-4). Similar, but more general,
is the idea of animation elements10, general object
hierarchies with a defined interface. However, both
approaches do not provide a general solution to this
problem.
Level of detail: Another problem of great consequence
is the reduction of scene complexity in interactive
applications. LOD concepts, such as progressive
meshes16, allow static objects to be fitted to the display
requirements. Recent techniques try to provide a viewdependent level of detail for static objects based on
mesh-simplification techniques or sometimes by
One of the main ideas of this paper is to structure the elements of an animated scene in terms of a linear space. This
idea is not new. Edelsbrunner11 appears to have first mentioned the term shape space and it was explicitly used by
Cheng, Edelsbrunner, and Fu6 for interpolating between
implicit descriptions of shapes. Alexa and Mller derive
spaces from any morphing technique1 and investigate conditions under which these spaces are linear. The idea of candidate spaces for object recognition was used by Turk and
Pentland25. The idea is to construct typical faces from a set
of photographs, which is done by computing the Eigenvectors of the space over all photographs (vectors of gray-levels). A space of three dimensional-shapes for recognition
and modeling of faces was used by Blanz and Vetter5.
In this work, we derive linear spaces by performing a
basis transformation for appropriately represented animated
polyhedral models. The resulting spaces can be used to
describe animation sequences in terms of principle components and their influence over time. The abstraction allows
for changes in the geometry as in the animation without
affecting each other. We can, for instance, use the same animation with different geometries. Morevoer, interesting
special effects may be achieved by exchanging single base
elements of the linear space. These and other applications
are demonstrated in Section 5.
3. The idea
In this paper, we restrict ourselves to the discussion of
scenes comprising animated polyhedral shapes described by
key-frame geometries B i . All polyhedral shapes are
assumed to have an isomorphic vertex-edge topology T .
Each shape B i is defined in terms of attributes of its boundary elements. Typically, at least a coordinate is associated to
a vertex, but other attributes such as a normal, a color value,
a texture coordinate, etc might be linked to it. Also, other
boundary elements may have attributes, e.g. normals might
be associated to faces rather than vertices. In general, vertex
attributes can be described by a vector of scalars and also
all vertices of a frame B i can be represented as a vector. We
assume that all base shapes have vectors of same length and
that the attributes are arranged identically. In the following,
Bi is used equivalently with the vector of these attributes.
The state of an object in a key-frame animation can then
be calculated by interpolating between two consecutive
key-frames. Formally, this can be described as:
A( t) =
i a i ( t ) B i
(1)
(2)
a i ( t ) B i .
(3)
(4)
principal component
bases
original key-frames
importance
factors
0
=
(5)
(6)
original/
all bases
50 bases
10 bases
5 bases
3 bases
Figure 3: Principal Component Analysis applied to the chicken animation. Prior to the SVD, base shapes are normalized.
The sequences above show frames 0, 80, 160, 240, 320, and 400 using different number of bases in the reconstruction.
encoding
size
ratio
original
14,544,000
14,548,800
1,818,600
1:8.3
364,800
1:39.8
181,860
1:79.97
109,116
1:133.28
1:0.99
By omitting a number of the unimportant base components, very high compression ratios can be achieved. Table
1 shows the compression ratios for the animation. However,
this animation is by far not typical. It comprises a highly
deformable object. For typical animation sequences, even
better results can be expected.
In the second example we applied the principal component representation in a facial animation system. This system uses several facial expressions coded as polyhedral
models with isomorphic vertex-edge topology. Key-frames
are generated by blending several expressions (e.g. saying
an A and smiling). In this system, the animation is represented as a vector over time that describes the linear combination of base shapes. Thus, animation representation and
geometry are already decoupled in this specific application.
In this example, we show how the geometry of the avatar
can be exchanged with another geometry making use of the
already existing animation descriptions.
Feature-based polyhedral morphing2,14,18 aims at finding
correspondences between different polyhedral models.
Most of these techniques rely on topological merging17, i.e.
a vertex edge topology is produced that contains both origi-
a)
b)
c)
d)
Figure 4: Exchanging geometry and in existing animations. a) A facial animation defined by a linear combination of base
shapes. b) A featured-guided morph between the original avatar mesh and a new mesh. Topological merging is used to produce a mesh which represents both shapes. Feature control assures that the same vertices represent common features (e.g.
mouth, eyes, etc.) c) The new mesh can be used with the existing animation with no additional user intervention. d) The
morph can even be applied while the animation is performed.
nal vertex-edge topologies as subgraphs. We have used a
feature-based morphing technique to produce a mesh that
can represent both the original avatar and another face. By
defining a few vertex-vertex correspondences, we make
sure that same vertices represent common features in the
source and the target model. This results in the convenient
fact that we can combine the new face with expressions of
the old one, e.g. we can add the smile defined in terms of
the original avatar to the new face. This means, by defining
the correspondences between the two neutral faces, we get
all other expressions of the face automatically. Play-back
animations authored for the original avatar can be directly
applied to the new face. In addition, we can morph between
the two faces while the animation is performed, since the
avatar and the new face now share the same vertex edge-
research and not yet applicable. Moreover, standard compression techniques were not exploited at all in this calculation. Additional compression can be achieved by
compressing the base shape matrices. Again, the principal
component analysis reorganizes the base shapes so that
bases with a higher index will contain more zeros. This
could be exploited with an additional entropy encoding.
The order of base objects naturally supports progressive
transmission of the animation base shapes. At the same
time, the inherent hierarchy could be used for LOD techniques. This could go hand in hand with a progressive
mesh. The importance ordering of base shapes additionally
eases mesh simplification since only the geometric features
of a few meshes have to be taken into account for simplification. In experiments we found that standard simplification techniques can be extended to handle more than one
mesh.
The animation itself is represented by small vectors if the
number of necessary base shapes is small. Note that the
number of necessary base shapes is bound by the number of
key frames. While the number of base shapes in our examples is much less than the number of key frames, it might be
useful to break very long animation sequences into pieces.
It would be interesting to exchange only parts of a base
while streaming an animation.
Having small vectors represent the animation allows for
streaming over virtually every network in real time. The
decoupling of animation and geometry enables managing
and changing animations, e.g. exchanging the animated
object according to the clients display capabilities. Mapping existing animations to a new object offers new ways of
authoring.
The necessary operations to play back animations
defined as linear combinations of a small set of base shapes
are now part of the proposals for animation in MPEG4.
Acknowledgements
This work would not exist without Herbert Edelsbrunner,
who introduced the first author to the idea of shape spaces.
Part of this work was done when the first author was visiting Tel-Aviv University supported by the Hermann
Minkowski - Minerva Center for Geometry. Thanks to
Daniel Cohen-Or, David Levin, and Craig Gotsman for
fruitful discussions.
We also want to thank Ulrike Spierling and Manfred Gaida
for providing the avatar models and Jed Lengyel for making
the chicken sequence public. The chicken character was
created by Andrew Glassner, Tom McClure, Scott Benza,
and Mark Van Langeveld. This short sequence of connectivity and vertex position data is distributed solely for the
purpose of comparison of geometry compression techniques.
The Eurographics Association and Blackwell Publishers 2000.
References
1.
2.
Marc Alexa. Merging Polyhedral Shapes with Scattered Features. Proceedings of Shape Modeling International '99, 1999, refined version in The Visual
Computer, 16, 1, pp. 26-37, 2000
3.
4.
5.
6.
7.
Daniel Cohen-Or, David Levin, and Amira Solomovici. Three dimensional distance field metamorphosis.
ACM Transactions on Graphics, 1998
8.
Daniel Cohen-Or, David Levin, Offir Remez. Progressive Compression of Arbitrary Triangular Meshes. Proceedings of Visualization 99, 1999
9.
Michael F. Deering. Geometry Compression. SIGGRAPH '95 Proceedings, pp. 13-20, 1995